Skip to content

AgentGuard v0.2.2 Release Verification

This versioned record documents the completed v0.2.2 production release. It is not a live badge and does not replace older commit-scoped validation artifacts.

Published Identity

Identity Value
Product AgentGuard
PyPI distribution agentguard-evals
Python import package agentguard
Console command agentguard
Version 0.2.2
Supported Python 3.9–3.12

The release used GitHub OIDC Trusted Publishing through the protected pypi environment. No PyPI token or password was used. PyPI generated digital attestations for the uploaded distributions.

Release Validation

  • Full test suite: 1,157 passed, 15 skipped.
  • Release-artifact validation: passed.
  • Isolated installed-wheel package smoke: passed.
  • Adversarial and showcase metrics checks: passed.
  • Ruff and diff checks: passed.
  • Fresh public installation used --no-cache-dir and the production PyPI index.
  • Installed metadata resolved as agentguard-evals==0.2.2.
  • import agentguard, agentguard --version, and agentguard --help passed outside the source checkout.
  • A network-free mock-agent evaluation passed with score 100/100.

The separately maintained validation summary remains a dated, commit-scoped coverage snapshot and was not rewritten for this documentation update.

Artifact Verification

The public files downloaded from PyPI were byte-identical to the exact validated distributions retained by the successful workflow:

File Workflow and public SHA-256
agentguard_evals-0.2.2-py3-none-any.whl 703e35376b94776318b8bbaf9fee91b78f97f9ae07d943b538a32a54d35997df
agentguard_evals-0.2.2.tar.gz ce9608259cabcdf7248c09f39e992dba1b36d2dd4138438f3a354dff2cf401b4

Both archives reported metadata name agentguard-evals, version 0.2.2, and Requires-Python: >=3.9. The wheel retained the agentguard import package and console entry point. Repository-only tests, docs, examples, workflows, caches, and generated runtime state were absent.

Historical And Scope Notes

AgentGuard v0.2.1 remains a valid GitHub-only release. PyPI rejected its original distribution identity before upload, so no v0.2.1 package was published. TestPyPI was intentionally unused because its relevant namespace is unrelated to this project.

Local-agent execution is not inherently sandboxed and should not be treated as a host security boundary. Docker-backed evaluations provide stronger isolation when configured, but AgentGuard does not claim universal containment or formal security certification.